Answers as JSON

Text for people, data for programs

A person can read the answer I think this comment is a complaint. It has the word "full". For a program, this text is hard to use. The program needs the parts of the answer: the label and the reason.

JSON gives each part a name. Lesson 11 of the course Python for data explains JSON. json.loads changes JSON text into a dictionary.

import json

text = '{"label": "complaint", "reason": "A full dock."}'
data = json.loads(text)
data["label"]   # "complaint"

Loose instructions

Answer in JSON. is not enough. With these instructions, the simulated model answers this:

Here is the JSON:
{"category": "complaint", "explanation": "It has the word \"waited\"."}

The answer has two problems.

• A sentence comes before the JSON. json.loads cannot read the text. It raises a json.JSONDecodeError, which is a kind of ValueError.
• The model chose the names of the keys. The next answer can have other names.

A real model makes the same two mistakes.

Strict instructions

Tell the model three things.

1. The answer is JSON, and only JSON.
2. The names of the keys.
3. The values that you allow, such as the labels.

Classify the comment as one of: praise, complaint, question,
other. Answer with only JSON. Use the keys "label" and "reason".

In the code, these instructions are one text on one line. The text has quotation marks inside. So put single quotation marks around it in Python.

A model can still add text

Clear instructions make extra text rare. They do not make it impossible. Now and then, a model adds a sentence, or it puts the JSON in a code block:

```json
{"label": "complaint", "reason": "It has the word \"full\"."}
```

A program can find the JSON in such an answer. It takes the part from the first { to the last }.

text = 'Here is the JSON:\n{"label": "praise"}'
start = text.find("{")     # 18
end = text.rfind("}")      # 36
text[start:end + 1]        # '{"label": "praise"}'

find and rfind give -1 when the text does not have the character. This method works for one JSON object in the answer. That is the usual case.

A setting that makes the service return JSON

Many real services have a setting for JSON answers. You give the service a description of the keys and of the type of each value. The service then makes sure that the answer is JSON in this form. Use this setting when your service has it.

The setting controls the form of the answer. It does not make the content correct. The label can still be wrong. Lesson 8 checks the content.

Your turn

main.py has a comment and the loose instructions. Do the tasks in order.